Databricks Open Sources Omnigent for Unified AI Agent Governance

Databricks has released Omnigent, an open-source meta-harness designed to unify the fragmented landscape of AI agent development. Released under the Apache 2.0 license, this new tool acts as a common interface that sits above various command-line agents and SDKs, including Claude Code, Codex, and Pi. By treating these individual harnesses as interchangeable components, Omnigent allows engineers to compose, govern, and share AI agent sessions through a single, cohesive platform.

A Unified Interface for Agent Orchestration

The architecture of Omnigent is built on the observation that while different harnesses may interact with models in unique ways, their user-facing interfaces—inputting messages and files and outputting text streams and tool calls—are fundamentally similar. Omnigent standardizes these interactions, enabling users to swap between different agents without rewriting code. The system consists of a runner that wraps agents in a sandboxed session with a uniform API, and a server that provides centralized policy management and collaboration features.
This design addresses the common challenge of managing multiple agents simultaneously. Instead of manually copying text between different coding agents, search tools, and documentation, users can coordinate several agents as interchangeable workers under one orchestrator. The platform supports a variety of environments, allowing a single session to remain in sync across terminal, web, and mobile interfaces.

Governance and Collaborative Capabilities

Omnigent introduces a robust control layer that moves beyond simple prompt-based guardrails. By implementing stateful, contextual policies at the meta-harness level, the system can enforce specific rules, such as pausing an agent once it reaches a set spending threshold or requiring human approval before executing sensitive commands like a git push following an npm package installation. These policies can be applied across three levels: server-wide, per-agent, and per-session.
Collaboration is similarly streamlined through the ability to share live agent sessions via URL. This feature allows teammates to observe an agent’s progress in real time, comment on files, co-drive the session, or fork the conversation. Security is managed through the Omnibox sandbox, which provides OS-level isolation and can transform network requests to keep sensitive information, such as GitHub tokens, hidden from the agent and injected only through an approved egress proxy.

Practical Implementation and Flexibility

The project ships with two example agents to demonstrate its capabilities. Polly serves as a multi-agent coding orchestrator that plans tasks and delegates work to sub-agents, while Debby functions as a brainstorming partner that allows two different models to debate topics side-by-side. Users can define custom agents and policies using simple YAML files, which specify prompts, harnesses, and tools.
While the project is currently in an alpha stage, it offers a flexible framework for developers to integrate various models and infrastructure. It supports multiple credential types, including first-party API keys, subscription-based access, and Databricks workspaces. By providing a consistent environment for composition and control, Omnigent aims to simplify the workflow for engineers managing complex, multi-agent systems.